Larkin

x402 answers "did they pay?" Larkin answers "should we let them?"

Authorization middleware for the x402 agent payment protocol. When an AI agent pays your API via x402, Larkin scores the wallet's trust before your handler runs and returns an Ed25519-signed receipt of the decision. Block, warn, or surcharge low-trust agents — your call.

TypeScript SDK

npm i @larkinsh/x402

import { preflight } from "@larkinsh/x402/next";

export const GET = preflight(handler, {
  apiKey: process.env.LARKIN_API_KEY!,
  minScore: 40,
  mode: "block",
});

Hono and Express adapters: @larkinsh/x402/hono and @larkinsh/x402/express.

Receipt verifier

npm i @larkinsh/verify

import { verify } from "@larkinsh/verify";

const result = verify(receipt, publicKeyBase64Url);
// → { valid: true, payload: {...} } | { valid: false, reason: "..." }

Verify any receipt Larkin has ever issued using only this library and our published public key — no call to our infrastructure required. Receipts remain independently verifiable forever.

CLI:

npx @larkinsh/verify receipt.json

Other packages

• @larkinsh/mcp — MCP server exposing wallet checks as a tool for Claude Code, Cursor, and other MCP-aware agents.
• larkin-x402 (PyPI) — Python SDK with the same shape as the TypeScript client.

Repo layout

packages/x402-larkin         — TypeScript SDK (@larkinsh/x402)
packages/x402-larkin-verify  — Receipt verifier (@larkinsh/verify)
packages/x402-larkin-mcp     — MCP server (@larkinsh/mcp)
packages/x402-larkin-py      — Python SDK (larkin-x402 on PyPI)
tests/x402-demo              — End-to-end SDK consumer example

Why Larkin exists

In January 2026, an AI agent called Lobstar Wilde lost $250,000 on a single transaction because no one checked its reputation. The x402 protocol solved machine-to-machine payment but left authorization as an exercise for the reader. Larkin is that exercise, finished.

Documentation

Full docs and API reference: https://docs.larkin.sh Marketing and signup: https://larkin.sh Issue tracker: https://github.com/larkin-dev/larkin-public-tmp/issues

Contributing

See CONTRIBUTING.md and SECURITY.md.

License

MIT. The hosted scoring service that produces the trust scores is a separate proprietary codebase. Every receipt these packages verify works forever, independently of that service — that's the trust model.